Do Affirmative Action Bans Lower
Minority College Enrollment and
Attainment?
Evidence from Statewide Bans
Ben Backes
ABSTRACT
Using institutional data on race-specific college enrollment and completion, I examine whether minority students were less likely to enroll in a
four-year public college or receive a degree following a statewide affirmative action ban. As in previous studies, I find that black and Hispanic enrollment dropped at the top institutions; however, there is little evidence
that overall black enrollment at public universities fell. Finally, despite evidence that fewer blacks and Hispanics graduated from college following a
ban, the effects on graduation rates are very noisy.
I. Introduction
Affirmative action remains a divisive subject in the United States.
Proponents deem it necessary to equalize opportunities available to different races,
citing racial disparities in educational attainment or earnings. Others attack affirmative action as a policy that perpetuates inequality and stereotypes by devaluing
the achievements of those who benefit from the policy, in addition to being unfair
to other groups. In response to court and citizen challenges, between 1997 and 2004,
six states—Texas, California, Washington, Florida, Georgia, and Michigan—banned
public institutions from using race when considering applications. Using institutional
data on race-specific college enrollment and completion, I examine whether minority
students were less likely to enroll in a four-year public college or receive a degree
following a statewide affirmative action ban.
Ben Backes is a graduate student in economics at the University of California, San Diego. He is grateful
to Kate Antonovics, Prashant Bharadwaj, Gordon Dahl, Marjorie Flavin, Valerie Ramey, Rick Sander,
two anonymous referees, and various seminar participants for helpful discussions and comments. The
data used in this article can be obtained beginning s October 2012 through September 2015 from the
author at [email protected].
[Submitted November 2010; accepted May 2011]
ISSN 022-166X E-ISSN 1548-8004 © 2012 by the Board of Regents of the University of Wisconsin System
T H E J O U R N A L O F H U M A N R E S O U R C E S • 47 • 2
436
The Journal of Human Resources
Previous studies of affirmative action bans have focused on application and enrollment, many using selected samples of universities or states, but little is known
about effects on the university system as a whole or on graduation. Card and Krueger
(2005) examine college application behavior in Texas and California and find no
evidence that highly qualified minority applicants reduced their rate of applying to
top state institutions in response to affirmative action bans. Furthermore, Antonovics
and Sander (2010) find no evidence that enrollment rates of minority applicants in
the UC system were lowered by the ban in California. Long (2007) examines the
impact of affirmative action bans on seven selective state universities and finds that
the elimination of race-based preferences led to a statistically significant increase in
minority underrepresentation (defined by comparing enrollment to the statewide racial makeup of the relevant age group) of five percentage points.1 Additionally, he
finds that top-x percent programs (which guarantee graduating seniors in the top x
percent of their class, where x varies by state, admission to a state university) offset
some of the decline in minority students at some institutions. Like this paper, Hinrichs (2011) uses institutional data to analyze the effect of affirmative action on the
minority share of enrollment. However, the enrollment analysis is not as detailed—
there is little attempt to estimate effects at various points of the university selectivity
distribution, or to assess whether minority students were absorbed by private, outof-state or two-year institutions. Although Hinrichs (2011) is the only paper I’m
aware of attempting to measure the effect on college completion, it uses demographic
data, which (as discussed later) is not ideal. As shown in this paper, institutional
data can provide more informative estimates of the impact on college completion.
The main way this analysis adds to the previous literature is that it includes a
comprehensive sample of institutions, rather than a chosen sample of the most selective universities. This allows for an assessment of whether fewer minority students
enrolled in college following a ban, rather than simply documenting what happened
at high-tier institutions. In addition, the paper attempts to assess the impact of the
bans on later outcomes; specifically, college completion. Although previous studies
have mainly focused on the effect of the bans on college application behavior, it is
college attainment that is ultimately of interest. In the sample, little more than half
of first-time students go on to graduate, and graduation rates are even smaller for
minority students. Because there is slippage between college application and enrollment, and between college enrollment and graduation, an analysis of college
completion, rather than simply enrollment, is needed to determine the welfare implications of the policy change.
This paper investigates college enrollment and completion at four-year public
universities in states that banned affirmative action. Estimates reveal little change in
the share of students enrolling in four-year public institutions who were black, and
a decrease in those who were Hispanic. Furthermore, there is evidence that the
average college graduate was less likely to be black or Hispanic after a ban, especially in the top decile of institutions. Additional analysis shows that it is not likely
1. There are many documented benefits of attending a selective university: Kane (1998), Loury and Garman
(1995), and Hoekstra (2009) find that graduates of more selective colleges have higher earnings, although
Bowen and Bok (1998) show that attending a more selective college improves other outcomes, such as
the likelihood of (not getting a) divorce.
Backes
that private institutions, two-year institutions or neighboring states absorbed additional minority students. All in all, although the effect sizes were modest, estimates
show that there were fewer black and Hispanic students graduating from four-year,
public universities following the bans, and those who did graduate tended to do so
from less prestigious universities.
II. Data set
The data for this analysis come from the Integrated Postsecondary
Education Data System (IPEDS), a panel of university-level data for public institutions from 1990 to 2009.2 Two measures will be used as outcome variables: first,
enrollment; and second, the number of students from each entering class who graduated. Data for each of these measures are available by year and racial group. The
enrollment data will be used to capture the effect of affirmative action on enrollment
of minority students at universities of different selectivity. The data on graduation
is only available for the entering classes of 1996 through 2003 and gives the number
of students from each entering class (by race) who attained a degree in six years or
less.
The main disadvantages of the data set are lack of information on postcollege
outcomes and on specific individuals, making an examination of the mechanisms of
the effects difficult.3
In order to measure the effects on the intended targets of the policy, the sample
is restricted to four-year, public institutions. This restriction could result in misleading results if there were large impacts on out-of-sample institutions, such as private
or two-year universities. In a later section, I will show that there is no evidence that
either type of institution offset the effects of the affirmative action bans.
III. Estimation Strategy and Overview of Data
Estimates are based on a comparison of the pre and post policy
change cohorts, controlling for general nationwide trends using time dummies. For
the share of race j at institution i located in state s at time t , a very simple specification can be written as
(1)
Enrollment sharejist = β0 + β1 • X st + θ • banst + γ i + γ t + ε ist
(2)
Graduate sharejist = β0 + β1 • X st + θ • banst + γ i + γ t + ε ist
where ban is equal to 1 if an affirmative action ban is in place, γ i denotes institution
fixed effects, X st represent time-varying state-level controls and γ t are year fixed
2. For a more complete discussion of the data used, see the Data Appendix.
3. Another way to measure the impact of the bans would be to use individual-level data to compare later
life outcomes of individuals who finished high school before the bans to those who finished after the bans.
However, using the ACS or any other data set not restricted to those likely to enroll in college is problematic
because college attainment rates of blacks and Hispanics are very low, making a change difficult to pick
up in the data.
437
438
The Journal of Human Resources
effects.4 θ is the measured impact of an affirmative action ban and represents the
expected change in the outcome variable when an institution becomes affected by a
statewide ban.5 Due to the heterogeneous nature of the treatment, the coefficients
should be interpreted as the average treatment effect across all treated institutions.
An alternative interpretation is that the model is a test of the null hypothesis that
each treatment has zero effect.
In this specification, the parameter of interest θ is identified if affirmative action
policies are uncorrelated with unobserved factors that affect the outcome variables,
such as time-varying statewide conditions. The vector X st is designed to control for
two factors that could potentially influence college attainment outcomes and be correlated with bans of affirmative action: first, statewide economic characteristics at
the time of application to college; and second, general statewide attitudes and policies toward education. For economic characteristics, I use yearly averages of the
state unemployment rate, fraction of individuals with a high school degree, fraction
of individuals with a college degree, and average income (these averages are taken
over individuals older than 25 years old). The education policy variables include
whether a state instituted a consequential accountability system prior to the relevant
year, whether an exam was required for graduating high school, and whether a topx percent program had been instituted.6 The sources for these variables are detailed
more thoroughly in the Appendix.
In addition, to assess how university selectivity affects the response to an affirmative action ban, institutions are divided into three SAT selectivity groups: the
decile of most selective universities, universities in the second and third deciles, and
the remainder.7 Selectivity groups are based on test scores of incoming students in
2007. The choice of 2007 would be problematic if the selectivity measure changed
in response to the treatment. However, the hierarchy of institutions is remarkably
constant over time (see Hoxby 2009), so using test scores from the pretreatment
period (if available) likely would not change the groupings.8 The reasoning behind
the number of group definitions is Kane’s (1998) finding that blacks and Hispanics
enjoy large admissions advantages at institutions whose mean SAT scores were in
the top fifth of all four-year schools. As a result, the initial analysis used regressions
for each of the top two deciles. However, there are few institutions in the second
decile in the policy change states, resulting in very large standard errors if these are
not grouped with the third decile. In any case, most of the effects of the policy
change occur in the top decile. Furthermore, I find little impact of the policies on
4. This can be thought of as a difference-in-differences approach with multiple policy-change years. The
ban term would be the interaction term in a traditional DD specification, with the level effects being
absorbed by the institution and year dummies.
5. Affirmative action bans could change high school effort levels. However, the effects of the bans appear
roughly constant over time, so the estimation results likely reflect changes in applications and admissions
rather than in precollege behavior (which presumably should have a time-varying component as students
are given more time to adjust effort).
6. None of the results in this paper are sensitive to including these controls.
7. I use test scores, rather than admissions rates, because admissions rates are not necessarily informative
about selectivity; see Hoxby (2009).
8. As an alternative measure of university rankings, I created a group of the top 50 US News public
schools. Results from the US News group are nearly identical to those in the top decile of test scores.
Backes
the institutions below the 70th percentile of SAT scores, and as a result combine
these institutions into one group. A small number of institutions reported neither
SAT nor ACT scores. Because these schools are observationally similar to low selectivity schools, I add the missing selectivity group schools to the low selectivity
category; this has no impact on results.
I define a state as being affected by an affirmative action ban when it is deemed
(by a court case, referendum, or law passed) that in-state public universities must
change their admissions procedures. As a result, I code Texans as first being affected
in 1997 because it was the first year Texan universities conducted admissions after
the conclusion of the Hopwood case. For Florida, I use 2001 because it was the first
year affirmative action was banned, even though the top 20 percent plan was introduced in 2000. The other states are more straightforward: California in 1998, Washington in 1999, Georgia in 2002, and Michigan in 2004.
Although this paper uses shares of enrollment and graduation, there are three
possible functional forms of the outcome variables: levels (for example, the number
of black students enrolling in a given year), logs, or shares of the total. Results will
be presented using race-specific shares as the outcome variable, although all three
specifications give very similar results. In each regression, universities are weighted
by their total enrollment or total graduates in 1996 (the final year in which no state
had adopted a ban), depending on the outcome variable. Reported standard errors
are robust to clustering at the state level.
A. Summary Statistics
Institutional summary statistics for four-year, public universities are presented in
Table 1. Numbers for enrollment and graduates are institutional averages for the
prepolicy change period: the three years before a ban in affected states and 1994–
96 for nonban states.
The first three columns show the entire sample of 526 institutions in all states,
followed by institutions in ban states and nonban states. About 20 percent of universities were in the six states affected by affirmative action bans. Universities in
ban states tend to be larger and have a higher proportion of Hispanic and Asian
students. Other measures, such as test scores and statewide economic characteristics,
are similar across the state groupings. The last three columns show summary statistics for the three selectivity groups for institutions located in ban states. The most
selective universities have students with higher test scores (by construction), have
much higher enrollment, and have a higher fraction of Asians and lower fraction of
blacks. Finally, a comparison of average enrollment and average graduates shows
that graduation rates increase as university selectivity increases.
B. Graphical Representation of the Effect of the Bans
A first look at enrollment over time in affected states is shown in Figure 1, which
shows enrollment shares of blacks and Hispanics aggregated by state and selectivity
group in a time window around each state’s ban year.9 The figure shows clear drops
9. Note that these figures do not use a common scale, because it would make changes in states with low
black or Hispanic populations very hard to detect.
439
Black graduates
percent
Hispanic graduates
percent
White graduates
percent
Asian graduates
percent
Black enrollment
percent
Hispanic enrollment
percent
White enrollment
percent
Asian enrollment
percent
Unknown
enrollment
percent
Graduates, all
Enrollment, all
Table 1
Summary Statistics
1848
(1454)
14.6
(21.7)
13.8
(17.3)
58.3
(27.3)
10.1
(13.2)
3.2
(4.0)
1035
(1189)
13.0
(22.4)
12.3
(17.0)
58.4
(28.2)
10.0
(13.3)
685
(856)
12.5
(23.4)
5.2
(10.5)
73.2
(27.5)
4.7
(8.9)
Ban States
1321
(1153)
14.2
(23.2)
5.9
(11.0)
72.6
(27.1)
4.7
(8.8)
1.8
(3.5)
All States
601
(732)
12.4
(23.7)
3.5
(7.3)
76.7
(26.2)
3.4
(6.9)
1195
(1031)
14.1
(23.6)
4.0
(7.7)
76.0
(25.9)
3.4
(6.8)
1.5
(3.2)
Nonban States
By State
476
(366)
16.7
(26.2)
14.9
(19.8)
53.8
(30.4)
8.0
(11.5)
1241
(752)
18.8
(25.1)
16.5
(20.1)
53.7
(29.6)
8.0
(11.1)
3.3
(4.4)
Low Selectivity
1743
(1239)
5.5
(4.2)
5.9
(4.9)
74.0
(16.1)
9.2
(12.0)
2750
(1655)
6.2
(3.8)
6.5
(5.3)
74.2
(14.4)
9.6
(12.0)
2.8
(3.2)
Medium Selectivity
By Selectivity
3070
(1407)
3.8
(2.4)
8.5
(4.6)
58.4
(22.2)
22.5
(18.7)
3835
(1699)
4.8
(2.7)
10.3
(5.8)
58.2
(20.7)
23.1
(19.1)
3.9
(3.2)
High Selectivity
440
The Journal of Human Resources
2.2
(4.6)
582
(63)
16,710
(1951)
0.20
(0.37)
0.87
(0.04)
0.25
(0.04)
0.04
(0.01)
526
3.7
(4.8)
589
(63)
17,986
(1503)
0.57
(0.43)
0.86
(0.04)
0.25
(0.02)
0.04
(0.01)
101
1.8
(4.5)
580
(63)
16,402
(1923)
0.11
(0.29)
0.88
(0.04)
0.25
(0.05)
0.04
(0.01)
425
69
3.8
(5.3)
553
(36)
20
2.9
(3.3)
633
(19)
12
4.3
(3.8)
702
(29)
Notes: 1990–2009 pooled IPEDS, public universities. Affected universities are those located in Texas, California, Washington, Florida, Georgia, and Michigan. Selectivity
groups defined according to reported 75th percentile SAT math score (see text). Selectivity group averages reported for institutions in ban states only. Additional summary
statistics available from author.
Number of institutions
Unemployment
Share over high
school
Share over college
School accountability
Unknown graduates
percent
75th percentile SAT
Math
Income (in dollars)
Backes
441
The Journal of Human Resources
Shares of Enrolment
442
Years Since Ban
Figure 1
Enrollment Share by State and Selectivity Group
in minority enrollment at each state’s most selective schools. The bottom graph of
control states is an aggregation of states adjacent to affirmative action banning
states.10 For these states, it appears there may have been a decrease in black enrollment at the most selective institutions following the bans.
Figure 1 underscores the importance of accounting for time trends, because some
states have steady preexisting trends in minority enrollment. A simple difference-indifference regression comparing highly selective universities in Florida to those in
nonban states would suggest that the affirmative action ban increased the share of
Hispanic students enrolling, because average enrollment was higher in Florida after
the ban and roughly constant in nonban states. Later in the paper I will give evidence
that adding a state-specific linear time trend removes the possible bias from these
increases over time.
Define the number of graduates from each entering class to be the number of
students from that class who would go on to graduate in six years or fewer. Figure
2 shows the shares of black and Hispanic graduates by enrollment year. Each point
on the graph shows, for a given enrollment class, what percentage of the cohort’s
10. The group of adjacent control states includes New Mexico, Oklahoma, Arkansas, Louisiana, Oregon,
Nevada, Arizona, Idaho, Alabama, North Carolina, South Carolina, Tennessee, Indiana, Ohio, Wisconsin,
and Illinois.
Shares of Graduates
Backes
Years Since Ban
Figure 2
Graduate Share by State and Selectivity Group
graduates were black or Hispanic.11 Figure 2 reveals similar patterns for graduation
as enrollment, although the changes are not as pronounced. For most states, there is
little evidence of an effect for the middle or low selectivity institutions for either
graduation or enrollment.
Table 2 shows average means of black and Hispanic enrollment for each of the
selectivity groups for the three years before and after each state’s ban. These basic
means show the same patterns in black enrollment as in Figure 1—falls at the most
selective institutions and little change in the other two groups. As shown in the
graphs, there were fewer blacks enrolled in the most selective universities in the
group of control states after the bans. However, the standard deviations for the group
of control institutions are large, which isn’t surprising as it represents an average of
a heterogeneous sample of states. For Hispanics, the simple sample averages don’t
show a clear patter due to states that experienced gradual rises in Hispanic enrollment
(such as Florida and Georgia), but some states, such as Texas and California, had
much lower Hispanic enrollment at their most selective institutions.
11. Another possible way of measuring the effect of affirmative action on the number of graduates would
be to measure the amount of BAs awarded by race and year. However, this method is not ideal because,
for any given year, it is not possible to tell when the students who graduated enrolled (and, thus, whether
they were affected by affirmative action bans or not). Finally, when using race-specific graduation rates,
the effects are very imprecisely estimated.
443
444
The Journal of Human Resources
Table 2
Sample Averages of Enrollment Share by State and Selectivity
Blacks
Texas
Preban
Postban
California
Preban
Postban
Washington
Preban
Postban
Florida
Preban
Postban
Georgia
Preban
Postban
Michigan
Preban
Hispanics
High
Medium
4.6
(0.59)
2.8
(0.13)
5.9
(3.6)
6.6
(5.2)
4.2
(2.5)
2.6
(1.3)
2.8
(1.3)
2.9
(1.3)
3
(0.3)
2.4
(.12)
2.5
(0.49)
2.9
(1)
9.9
(2.1)
8.5
(1.6)
9.7
(2.2)
10
(2.1)
43
(44)
43
(43)
11
(1.2)
12
(0.88)
5.4
(.42)
5.5
(0.17)
6.2
(2.9)
6.2
(3.7)
28
(24)
29
(25)
2.4
(0.62)
3.1
(0.58)
1.7
(0.45)
1.9
(0.32)
1.8
(1.1)
2.1
(1.1)
13
(9.5)
14
(12)
4.3
(1.8)
4.3
(1.5)
2.6
(.81)
2.7
(0.63)
1.9
(.68)
2.1
(0.75)
18
(25)
21
(26)
3.8
(2.6)
4
(2.5)
3.6
(4.2)
3.3
(3.9)
3.6
(7.5)
3.5
(6.8)
7.3
7.6
(2.6)
(3.2)
Postban
5.5
8.1
(2)
(2.7)
Adjacent state control group
Preban
7.5
6.8
(3.4)
(5.8)
Postban
6.6
7
(2.9)
(5.9)
Low
High
Medium
14
(1.3)
11
(1.9)
9.4
(1.8)
8.4
(1.1)
24
(27)
24
(27)
9.5
(5.9)
7.8
(5.7)
15
(3.6)
10
(1.6)
13
(4.4)
14
(4.9)
28
(12)
25
(12)
2.6
(1.1)
2
(0.63)
4.5
(0.66)
3
(0.78)
3.6
(0.46)
3.3
(0.55)
20
(29)
22
(30)
8.8
(2.2)
11
(2)
Low
4.1
(0.83)
4.3
(0.68)
23
(27)
24
(27)
Notes: 1990–2009 pooled IPEDS, public universities. Averages taken over the three years before and after
the passage of an affirmative action ban. Selectivity groups defined by SAT scores of incoming students
(see text). Institutions weighted by total enrollment in 1996.
Backes
IV. Estimation Results
A central question concerns the effect of affirmative action bans on
the enrollment shares of minority students at universities of different levels of selectivity. Results for different specifications of Equation 1 are shown in Table 3 as
a way to test for sensitivity. Specifications tested include changing the nature of the
fixed time trends (linear state trends, linear university trends, or squared university
trends), restricting the sample to include a narrower window around the bans and
restricting the nonban state sample to only states adjacent to a banning state. Reported coefficients are expressed in percentage terms: a coefficient of one represents
a one percentage point change.
The first row of regressions shows the estimated effect of affirmative action bans
on the black share of total enrollment for the pooled sample of all universities for
the various specifications. Other than the first column, which does not include any
time trends, the point estimates across specifications are similar. The coefficients are
small (around 0.4) and insignificant or marginally significant, suggesting that if affirmative action lowered the share of enrolling students in four-year public institutions who are black, the effects were modest.12 The specifications that include some
sort of fixed time trend—whether state- or university- specific—give very similar
results: Columns 2–4 and 7 are nearly identical. Columns 5 and 6, which either
restrict the years included or the set of control states, give similar estimates but tend
to have larger standard errors, which is not surprising because the sample is
smaller.13
The next row restricts the sample to the decile of the most selective universities.
Again, as long as a time trend is included, results are very similar across specifications. Results show a highly significant drop of about 1.6 percentage points in the
number of enrolling students who are black, consistent with previous studies of
affirmative action bans at selective universities.
For institutions in the medium-selectivity group (the second and third deciles of
SAT scores), there is a marginally significant increase of about 0.4 in the percentage
of incoming students who are black. This number is much smaller in magnitude
than the decrease in enrollment at the top tier of institutions, so it appears that the
response of the medium-selectivity institutions was less pronounced than at the top—
the capacity of medium-selectivity schools did not expand to fully offset the drops
at high selectivity schools, even after taking into account the greater number of
schools in the medium group.14 Finally, there is no evidence that the remaining low
selectivity institutions experienced a change in black enrollment following the bans.
12. When the outcome variable used is the level of black enrollment or its log, the coefficient is not
significant.
13. The restricted years include 1995–2001, designed to capture a period in which many states changed
their affirmative action laws.
14. To compare effects across selectivity groups, coefficients should be weighted by the total number of
students in each group. Using the Table 1 to perform a rough calculation yields about 46,000 (12 schools
* 3,835 students/school) students enrolling in the high group and 55,000 (20*2,750) students enrolling in
the medium group each year in affected states.
445
−0.18
(0.60)
364
0.67
0.36*
(0.15)
116
0.61
−1.49***
(0.35)
46
0.52
−0.33
(0.26)
526
0.65
University
squared
trends
(4)
−0.48
(0.79)
364
0.049
0.68***
(0.17)
116
0.047
−1.52***
(0.40)
46
0.18
−0.45
(0.27)
526
0.022
Years
restricted
(5)
−0.69
(0.57)
186
0.14
0.61
(0.46)
60
0.044
−1.08*
(0.43)
19
0.18
−0.45
(0.30)
265
0.045
Adjacent
states
(6)
−0.11
(0.35)
186
0.2
0.41
(0.26)
60
0.2
−1.69***
(0.25)
19
0.53
−0.36*
(0.17)
265
0.089
Adjacent
and trend
(7)
Notes: 1990–2009 pooled IPEDS, public universities. Selectivity groups defined according to reported 75th percentile SAT math score (see text). Controls include average
state income, unemployment, share of adults with college and high school degrees, whether a state adopted school accountability, and whether a state adopted a top-x
percent plan (see text). Universities weighted by total enrollment in 1996. Standard errors robust to clustering at the state level. In addition to the controls listed above,
all models include year dummies, while Column 2 adds state-specific linear time trends, Column 3 university-specific linear time trends, Column 4 university-specific
quadratic time trends, Column 5 a restricted time horizon, Column 6 a restricted group of control states, and Column 7 a restricted group of states with state-specific
linear time trends (see text). * Significant at 10 percent, ** 5 percent, *** 1 percent.
Number
R-squared
−0.18
(0.50)
364
0.57
0.19
(0.61)
364
0.089
Low-selectivity institutions
Number
R-squared
−0.17
(0.49)
364
0.19
0.61
(0.34)
116
0.035
Medium-selectivity institutions
Number
R-squared
−1.65***
(0.25)
46
0.44
0.42*
(0.18)
116
0.43
−1.65***
(0.25)
46
0.35
−1.01**
(0.32)
46
0.1
High-selectivity institutions
−0.38
(0.19)
526
0.53
University
trends
(3)
0.42*
(0.17)
116
0.28
−0.38*
(0.19)
526
0.11
−0.03
(0.34)
526
0.031
All institutions
Number
R-squared
State
trends
(2)
Year
dummies
(1)
Table 3
Effect of Ban on Black Enrollment Share Under Various Specifications
446
The Journal of Human Resources
Backes
To summarize, across the specifications, the predicted fall in black enrollment at
the top universities following an affirmative action ban is about 1.6 percentage points
of enrollment. Additionally, there were possible increases in the second tier of institutions and very small, nonsignificant decreases in the bottom 70 percent of universities. Because results are similar regardless of the specification of the time trend,
I will take the simple state trends (Column 2) to be the preferred specification and
use them for the rest of the paper.
Table 4 shows a summary of regressions using the specification used in Table 3,
Column 2 for various samples. Thus, the first panel of Column 1 is identical to Table
3’s Column 2. Column 1, Panel 2 shows the same regressions for the share of
Hispanic enrollment. Coefficients are qualitatively similar to those for blacks but
consistently larger in magnitude, which is not surprising due to the higher initial
levels of Hispanic enrollment in banning states. The first row shows that at the
average institution, the share of Hispanic enrollment fell by about 1.4 percentage
points, although this is only significant at the 10 percent level. Furthermore, the
enrollment drops at the top selectivity group was larger for Hispanics than for
blacks—about 2.9 percentage points. There were no significant changes in the second
tier and nonsignificant, though somewhat large, falls at the lowest tier.
For selective institutions, the regression coefficients represent large changes in the
outcome variable. For black students at the most selective universities, average enrollment share in the post policy change period was 4 percent (author’s calculation).
According to the regression results, black enrollment share at a given university
would have been 1.6 percentage points higher had affirmative action not been
banned. Thus, the bans led black enrollment to be 1.6/(4 + 1.6) = 29 percent lower
at top institutions than it would have been in the absence of a policy change. At the
average institution, the total number of black students fell by about 0.38/11.7 percent—a relatively small change that is only marginally significant in some specifications. For Hispanics, enrollment share at the top institutions fell by 2.9 percentage
points. Compared to the postpolicy change mean of 11.3 percent (author’s calculation), this is a change of about 20 percent. Note that in absolute terms the number
of Hispanics at the top institutions fell by more than for blacks; however, because
there are many more Hispanics, the percentage change in Hispanic enrollment is
smaller. For Hispanics at the average institution, the total number of students fell by
about 8 percent.15 In other words, it does appear that the total enrollment of Hispanics in four-year public institutions was reduced by affirmative action bans (and
possibly a very slight reduction for blacks as well).
Because each of the coefficients represents the percentage of total enrollment, it
is important to note that total enrollment was unaffected by the policy change, as
shown in Panel 3, Column 1 of Table 4. Coefficients reported are from regressions
with log total enrollment as the outcome variable. The coefficients are consistently
small and insignificant: the policy change appears not to affect total enrollment at
the average institution or at the top tier.
15. For both blacks and Hispanics, using log enrollment as the outcome variable yields coefficients of
similar magnitude to these rough calculations of percentage change in enrollment, so these results are
robust to changes in how the outcome variable is measured.
447
448
The Journal of Human Resources
Table 4
Effect of Ban on Black and Hispanic Enrollment and Graduate Shares
Enroll
public
(1)
Graduate
public
(2)
Enroll
private
(3)
Enroll
two-year
(4)
Number
−0.38*
(0.19)
526
−0.62***
(0.16)
520
−0.52
(0.28)
1029
0.01
(0.13)
983
High-selectivity
institutions
Number
−1.65***
(0.25)
46
−1.24**
(0.44)
46
Medium-selectivity
institutions
Number
0.42*
(0.17)
116
0.37
(0.24)
116
−0.69
(0.61)
74
Low-selectivity
institutions
Number
−0.17
(0.49)
364
−0.54
(0.35)
358
−0.65*
(0.28)
890
Number
−1.36*
(0.65)
526
−0.59
(0.36)
520
−0.72***
(0.20)
1029
High-selectivity
institutions
Number
−2.87***
(0.55)
46
−1.81**
(0.59)
46
−1.33
(1.00)
65
Medium selectivity
institutions
Number
−0.54
(0.38)
116
0.61
(0.64)
116
−0.52
(0.44)
74
Low-selectivity
institutions
Number
−1.00
(1.02)
364
−0.03
(0.20)
358
−0.66**
(0.22)
890
Panel 1: Blacks
All institutions
Panel 2: Hispanics
All institutions
Panel 3: Log total
All
−0.014
(0.02)
0.17
(0.49)
65
−0.32
(0.38)
983
0.024**
(0.01)
Note: see notes from Table 3.
Next, I briefly turn to white and Asian enrollment. One drawback to the regressions for Asians and whites is that there was an increase in students who did not
report their race, and the majority of these students are likely white or Asian (dis-
Backes
Table 5
Effect of Ban on White and Asian Enrollment and Graduate Shares
Panel 1: Whites
All institutions
Number
High-selectivity institutions
Number
Medium-selectivity institutions
Number
Low-selectivity institutions
Number
Panel 2: Asians
All institutions
Number
High-selectivity institutions
Number
Medium-selectivity institutions
Number
Low-selectivity institutions
Number
Enroll public
(1)
Graduate public
(2)
0.91*
(0.38)
526
−0.49
(1.56)
520
3.16**
(0.95)
46
−0.33
(3.02)
46
−0.48
(0.58)
116
−2.72
(2.20)
116
0.61
(0.90)
309
0.28
(0.66)
306
−0.23
(0.16)
526
0.11
(0.33)
520
−0.15
(0.41)
46
0.78
(0.86)
46
−0.32
(0.25)
116
−0.77*
(0.34)
116
−0.29*
(0.14)
309
0.03
(0.10)
306
Note: see notes from Table 3.
cussed in a later section). These results should be treated with caution and are not
the emphasis of this paper. Effects on white enrollment are shown in Table 5, Column 1, Panel 1. The results reveal that the total share of white students enrolling
increased by about 0.9 percentage points and that there were large increases in
enrollment—more than three percentage points—at the top institutions.
Panel 2 shows that the bans had little impact on the Asian share of enrollment at
the more selective institutions. When state trends are not included (regressions not
449
450
The Journal of Human Resources
shown), affirmative action bans are predicted to have a large positive impact on
Asian enrollment, but this result completely disappears once the time trends are
added. This is not surprising as many top-tier institutions, especially in California,
have experienced persistent rises in Asian enrollment. Once these fixed time trends
are accounted for, a ban does not predict any deviation in Asian enrollment from
this trend.
A. Graduation
To see how the effects on enrollment shares translate into graduate shares, I regress
race-specific graduate shares on whether an affirmative action ban is in place, in the
same manner as above. Recall that data are only available for the entering classes
between 1996 and 2003, so the panel is shorter than for enrollment. Column 2 of
Table 4 displays results for the shares of black and Hispanic graduates at public
universities.
The first row of Panel 1 shows a decrease in the total share of black college
graduates following a ban, mostly coming from the reductions at the most selective
institutions. The average institution’s share of black graduates fell by about 0.6
percentage points following a ban. As with enrollment, the effects are largest at the
most selective universities. The second row of estimates shows a 1.2 percentage
point decrease in the share of graduates who are black at the most selective institutions. For the medium-selectivity institutions, estimates are similar to enrollment—
a 0.37 percentage point increase. Finally, the least selective institutions show a nonsignificant decrease in graduate shares.
For Hispanics, Column 2 shows a nonsignificant drop in the share of graduates
at the average institution of about 0.6, smaller than the drop in enrollment in Column
1 and similar to overall fall in the share of black graduates. The share of Hispanic
graduates at the top decile fell by about 1.9 percentage points.
When comparing the enrollment and graduation, if lower quality students were
the ones displaced by the bans, one would expect the effects on graduates to be less
pronounced than on enrollment, because these students would have been less likely
to graduate. Looking at the three selectivity groups for blacks and Hispanics shows
this to be the case for Hispanics in each group and for blacks in all groups except
the lowest tier, where enrollment fell by −0.18 and graduation fell by −0.55 percentage points. The changes in this group presumably lead to the graduation effect
for blacks being larger than the enrollment effect for the entire sample (−0.63 and
−0.38), but the limitations of institutional-level data make it difficult to say more,
especially because the coefficients for the lowest tier are not very precisely estimated.
To summarize, the previous tables reveal a mostly clear picture of the effect of
the affirmative action bans. Confirming previous findings, there were large drops in
the black and Hispanic share of students enrolling and graduating from the top tier
of institutions. However, at the average institution, the drops in the black share of
enrollment were very small, with little evidence of change in the overall number of
blacks enrolling in college following the affirmative action bans. For the black share
of university graduates, the marginal effects at the average university are significant
and negative, although not large in magnitude, so there may have been a drop in
the likelihood of blacks graduating from college. At the average institution, the
Backes
Hispanic share of both enrollment and graduates dropped, although the coefficient
for graduates is not significant. Overall, the results for blacks and Hispanics are
generally similar, with both groups being pushed out of the most selective institutions
but having smaller changes for the university system as a whole.
B. Private Institutions, Two-Year Institutions, Out-of-State Students, and
Nonrace Reporters
One way the effects of affirmative action could have been mitigated would be if
private universities absorbed some of the students who would have otherwise attended public school. The previous analysis using institutional data suggests that
black and Hispanic enrollment fell at the most selective universities. If the number
of black and Hispanic students at the most selective private institutions rose following a ban, the implications would be quite different—the top students would have
been absorbed by the best private institutions, rather than being forced to attend
lower-tier public institutions. Results for a sample consisting of private institutions
are reported in Column 3 of Table 4. For blacks, the coefficient for the top selectivity
group is very small. For Hispanics, results are somewhat mixed, with some negative,
significant coefficients. In any case, there certainly does not appear to be an increase
in the share of minority private school enrollment—either overall or at the most
selective institutions—following the bans.
Another possible effect of the ban could have been to push minority students from
four-year to two-year institutions. Previous results showed a decrease in Hispanic
enrollment at the average four-year university following a ban. Column 4 of Table
4 shows enrollment results for the sample of two-year institutions. Institutions are
not split into selectivity groups because very few schools report SAT scores, so only
regressions for the average institution are shown. For blacks, there is no evidence
of any change in two-year enrollment following the bans. For Hispanics, there is a
small, nonsignificant decrease in enrollment share.
Another factor that could affect the welfare implications would be induced interstate migration—perhaps black and Hispanic students reacted to the bans by attending institutions in other states. Unfortunately, IPEDS does not report out-of-state
enrollment by race, so I use two somewhat indirect methods of testing. One test is
to regress minority share of enrollment on whether an adjacent state has enacted an
affirmative action ban. The basic idea of this regression is to see whether, for example, universities in Oregon had an increase in the minority share of enrollment
after 1998, the year California enacted its ban. Coefficients from these regressions
(not reported) are small and not significant—evidence that blacks and Hispanics did
not react to the bans by attending college in nearby states. The second test is seeing
whether states adjacent to banning states experienced increases in the total share of
out-of-state students. Again, regression results indicate that this is not the case.
Finally, one problem with the IPEDS data is that there are a substantial number
of students whose race is not reported. If the failure to report race is affected by the
bans, this could bias results. For example, in 1998 there was a sharp increase in the
number of students who did not report race at institutions in the University of California system. If the problem only occurred in the first year of a ban (as appears to
be the case in the California universities), then dropping the first year of a ban would
451
452
The Journal of Human Resources
provide a check on the results. When dropping the initial ban year, results for blacks
and Hispanics are very similar to the full sample, suggesting that it was individuals
of other races who changed their likelihood of reporting race following the bans.
Furthermore, Antonovics and Sander (2010) provide convincing evidence that, in
California, the rise in unknowns is due to an increase in whites and Asians not
reporting race (and in fact classify unknowns as white/Asian in their paper). If these
unknowns are indeed white or Asian, then drawing conclusions about what happened
to the enrollment shares of whites and Asians would be faulty. However, the evidence presented about blacks and Hispanics would still be valid.
C. Model Specification Checks
An example of an endogeneity problem would be if the bans were passed in response
to rising minority enrollment in public universities. If the bans were as good as
randomly assigned, conditional on the controls, they should not be able to predict
changes in enrollment in the time period leading up to the policy changes. A way
to test for this potential problem would be to add leads of the policy change variable
to the regressions to investigate whether these leads can predict changes in the
outcome variable (often referred to as a &apos+acebo’ test in the literature). However, using the one-year lead of the policy change is not informative because some
universities implemented the policy change in the year before they were forced to,
making nearly every model fail the placebo test.16 As a result, I run regressions with
a dummy for whether a state will change its affirmative action policy in either two
or three years in the future. Significant coefficients of this policy change lead would
indicate that changes in the racial composition of enrollment were associated with
the introduction of the bans. Results are presented in Table 6. For blacks, none of
the specifications are significant for any of the selectivity groups, so it appears
unlikely that the bans were passed in response to changes in black enrollment (or
something correlated with black enrollment). However, for Hispanics, some coefficients are significant and positive, although when using state or university time
trends, the coefficients lose significance. This could be interpreted as evidence that
adding either state or university specific fixed time trends is sufficient in eliminating
bias due to time trends in the data. However, it could also be evidence that states
instituted bans in response to growing Hispanic enrollment in public universities.
Results are robust to changes in the functional form of the amount of time a ban
has been in place (for example, adding a linear term in the number of years of
exposure to an affirmative action ban)—allowing for more flexibility does not result
in meaningful changes to results. Other changes in model specification that do not
affect the coefficients are whether the state-specific economic conditions and accountability measures are included. Finally, rerunning regressions although dropping
one state at a time, results (not shown but available from author) remain similar.
16. One way of dealing with these prepolicy change decreases in enrollment would be to switch the policy
change year to, for example, 1996 in Texas (rather than 1997); this does not affect the main results of the
paper. Another way of running these placebo tests would be to only include only Washington and California, the nonearly adopters. Results are similar to those discussed in this section.
0.96**
(0.35)
0.49
(0.32)
0.73***
(0.17)
1.45
(0.84)
0.04
(0.20)
0.50
(0.35)
0.04
(0.23)
−0.02
(0.36)
Year dummies
(1)
0.37
(0.23)
0.08
(0.25)
0.09
(0.17)
0.83
(0.48)
−0.17
(0.15)
−0.01
(0.24)
−0.12
(0.19)
−0.20
(0.31)
State trends
(2)
0.36
(0.23)
0.08
(0.25)
0.09
(0.18)
0.83
(0.50)
−0.17
(0.16)
−0.01
(0.24)
−0.12
(0.20)
−0.21
(0.32)
University
trends
(3)
0.46*
(0.22)
0.45
(0.25)
0.36
(0.19)
0.58
(0.42)
−0.13
(0.13)
0.04
(0.25)
−0.07
(0.16)
−0.29
(0.25)
University
squared trends
(4)
0.76*
(0.29)
0.69*
(0.30)
0.44
(0.26)
0.56
(0.29)
0.03
(0.18)
−0.18
(0.20)
−0.01
(0.18)
0.06
(0.36)
Years restricted
(5)
1.00*
(0.35)
0.86*
(0.33)
0.73***
(0.18)
1.42
(0.82)
−0.30
(0.14)
0.38
(0.35)
−0.03
(0.26)
−0.51
(0.29)
Adjacent states
(6)
Notes: 1990–2009 pooled IPEDS, public universities. Reported coefficients measure the average “effect” of the second and third leads of the policy change. See notes
from Table 3.
Low-selectivity
Medium-selectivity
High-selectivity
Panel 2: Hispanics
All
Low-selectivity
Medium-selectivity
High-selectivity
Panel 1: Blacks
All
Selectivity group
Table 6
“Effect” of Lead of Ban on Enrollment Share
Backes
453
454
The Journal of Human Resources
V. Discussion
This paper investigates college enrollment and completion at fouryear public universities in states that banned affirmative action. Estimates show that
the falls in black and Hispanic enrollment were confined to the decile of most
selective institutions. However, there is some evidence that fewer black and Hispanic
students graduated from college following the bans, although the effect sizes are
modest.
The results from the analysis suggest that affirmative action did succeed in raising
the shares of black and Hispanic enrollment at the top institutions but did not lead
to more blacks enrolling in college. However, the effects of affirmative action—both
at top-tier schools and the university system generally—are small relative to the total
population of minority students.
Appendix
Data
Institutional data is taken from the Integrated Postsecondary Education Data System (IPEDS), a survey of higher education institutions conducted by
the National Center for Education Statistics (NCES). Data was accessed through the
Data Center portion of the NCES web page. Every institution participating in a
federal student aid program (including Pell grants and federal student loans) must
report data on enrollment, degrees awarded, tuition and many other variables. My
main focuses are fall enrollment of full-time, first-time students and the number of
graduates from each enrolling class. The sample consists of enrollment data from
1990–2009, with the exception of 1999, which is not posted on the IPEDS web
page. I include four-year public universities that report data for each year in the
sample.17 To construct a measure of institution selectivity, I use two reports from
2007: the 75th percentile math SAT score and75th percentile ACT score. First, I
divide schools into both SAT and ACT deciles. Schools in the top SAT group are
then assigned to the group of most selective institutions. Universities in the top decile
of ACT scores that do not report SAT scores are also assigned to the top selectivity
group. The process is repeated for the second and third deciles of test scores, which
make up the second selectivity group used in the paper. The final sample consists
of a panel of 526 institutions.
Statewide School Accountability
A concurrent factor that could influence educational outcomes was the introduction
of state school accountability systems, which have been found (for example, Hanushek and Raymond 2005) to boost student achievement. These could be especially
problematic if there was a disproportionate effect on minority students. To control
for the introduction of these systems, I add dummies for whether an insti17. About 80 percent of the institutions that reported data in 1990 meet these criteria.
Backes
tution’s state introduced consequential school accountability, as shown in Table 2 of
Miller and Zhang (2006). Following Hanushek and Raymond (2005), I define a
consequential school accountability system to be one where school test performance
is publicly disseminated and there are consequences for the results of these reports.
I also add dummies for whether a state requires the passage of an exam to graduate
from high school (taken from the Education Commission of the States web page).
Top-x Percent Programs
Three of the affirmative action banning states enacted a top-x percent rule after the
ban. As mentioned earlier, Long (2004) and Long (2007) show that a top-x percent
rule fails to replace affirmative action in maintaining the share of minority students
at top universities. Thus if minorities are affected by the elimination of race-based
preferences, the estimation model would still show a net effect, even without controlling for top-x percent policy changes. The inclusion of these terms does not have
a large impact on the estimates.
References
Antonovics, Kate, and Richard H. Sander. 2010. “Affirmative Action Bans and the ‘Chilling
Effect’.” Working paper.
Bowen, William G., and Derek C. Bok. 1998. The Shape of the River: Long-term
Consequences of Considering Race in College and University Admissions. Princeton:
Princeton University Press.
Card, David, and Alan B. Krueger. 2005. “Would the Elimination of Affirmative Action
Affect Highly Qualified Minority Applicants? Evidence from California and Texas.”
Industrial and Labor Relations Review 58(3):416–34.
Hanushek, Eric A., and Margaret E. Raymond. 2005. “Does School Accountability Lead to
Improved Student Performance?” Journal of Policy Analysis and Management 24(2):297–
327.
Hinrichs, Peter. 2011. “The Effects of Affirmative Action Bans on College Enrollment,
Educational Attainment, and the Demographic Composition of Universities.” Review of
Economics and Statistics. Forthcoming.
Hoekstra, Mark. 2009. The Effect of Attending the Flagship State University on Earnings:
A Discontinuity-Based Approach. Review of Economics and Statistics 91(4):717–24.
Hoxby, Caroline M. 2009. “The Changing Selectivity of American Colleges.” Journal of
Economic Perspectives 23(4):95–118.
Kane, Thomas J. 1998. “Racial and Ethnic Preference in College Admissions.” In The
Black-White Test Score Gap, ed. Christopher Jencks and Meredith Phillips, 431–56.
Washington, D.C.: Brookings Institution.
Long, Mark C. 2004. “Race and College Admissions: An Alternative to Affirmative
Action?” Review of Economics and Statistics 86(4):1020–33.
———. 2007. “Affirmative Action and its Alternatives in Public Universities: What Do We
Know?” Public Administration Review 67(2):315–30.
Loury, Linda D., and David Garman. 1995. “College Selectivity and Earnings.” Journal of
Labor Economics 13(2):289–308.
Miller, Amalia R., and Lei Zhang. 2006. “Did Welfare Reform Improve the Academic
Performance of Children in Low-Income Households?” Department of Economics,
University of Virginia. Unpublished.
455